The WorkChapter 05 / 05

Chapter 05 / 2026 / Services

Finance Agent

A personal finance app that runs on your own Mac. It pulls in your bank transactions, sorts them into categories and shows where the money goes each month. You can ask it questions in plain English, and every number in its answer comes from your own records.

What it is

Bank data is too private for the cloud

My bank data shouldn't have to leave my laptop.

A personal finance app that runs on your own Mac.
The app's list of transactions for a made-up household, each with its shop, its amount and the category it was given.
Who it is forPeople who want help with their money without sending their bank data to someone else's server.

How it works

Local, encrypted, read-only

The model can ask the ledger, but it never makes up a number.

1. Bring in the transactionsThe ledger kept on your own desk, under lock: transactions come in from your banks and stay encrypted on the Mac.
2. Sort themThe category each transaction was given, by a rule, a correction you made or the local model.
3. Show the monthThe month as a ledger page: what each category took, added up into the month's totals.
4. Answer questionsA question in plain English, its answer, and the ledger query that produced the figure in it.
The app's list of transactions for a made-up household, each with its shop, its amount and the category it was given.

What went wrong

The eight-second summary

Eight seconds to open a month. That had to go.

On a real ledger the monthly summary took 8.3 seconds, and every click on a month waited for it.
Profiling showed three-quarters of that time went to loading full account details only to read their ids.

What it measured

Fast, and the numbers add up

Fast, and every figure adds up.

8 s → <300 msThe month's overview now appears in under a third of a second instead of eight.
150+More than 150 separate changes, each written, checked and merged on its own.
6The demo shows six made-up households, so no real bank data is ever on show.

My part and the stack

My partBuilt it himself, end to end: the encrypted ledger, the categoriser, the Assistant, the app, the landing page and the public demo on synthetic data.
A question asked in plain English and its answer, with the ledger query that produced the figure shown beneath it.